Re: [PATCH v2] perf: libdw support for powerpc [ping]

On Tuesday, June 13, 2017 5:55:09 PM CEST Ravi Bangoria wrote:

Hi Mark,

On Tuesday 13 June 2017 05:14 PM, Mark Wielaard wrote:

quoted

I see the same on very short runs. But when doing a slightly longer run,
even just using ls -lahR, which does some more work, then I do see user
backtraces. They are still missing for some of the early samples though.
It is as if there is a stack/memory address mismatch when the probe is
"too early" in ld.so.

Could you do a test run on some program that does some more work to see
if you never get any user stack traces, or if you only not get them for
some specific probes?

Thanks for checking. I tried a proper workload this time, but I still
don't see any userspace callchain getting unwound.

  $ ./perf record --call-graph=dwarf -- zip -q -r temp.zip .
  [ perf record: Woken up 2891 times to write data ]
  [ perf record: Captured and wrote 723.290 MB perf.data (87934 samples) ]


With libdw:

 $ LD_LIBRARY_PATH=/home/ravi/elfutils-git/usr/local/lib:\
    /home/ravi/elfutils-git/usr/local/lib/elfutils/:$LD_LIBRARY_PATH\
    ./perf script

  zip 16699  6857.354633:      37371 cycles:u:
                   ecedc xmon_core
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux) 8c4fc
__hash_page_64K (/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux)
83450 hash_preload
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux) 7cc34
update_mmu_cache (/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux)
330064 alloc_set_pte
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux) 330efc do_fault
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux) 334580
__handle_mm_fault (/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux)
335040 handle_mm_fault
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux) 7bf94
do_page_fault (/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux)
7bec4 do_page_fault
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux) 7be78
do_page_fault (/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux)
1a4f8 handle_page_fault
(/usr/lib/debug/lib/modules/4.11.0-3.el7.ppc64le/vmlinux)

  zip 16699  6857.354663:     300677 cycles:u:

  zip 16699  6857.354895:     584131 cycles:u:

  zip 16699  6857.355312:     589687 cycles:u:

  zip 16699  6857.355606:     560142 cycles:u:

Just a quick question: Have you guys applied my recent patch:

commit 5ea0416f51cc93436bbe497c62ab49fd9cb245b6
Author: Milian Wolff [off-list ref]
Date:   Thu Jun 1 23:00:21 2017 +0200

    perf report: Include partial stacks unwound with libdw
    
    So far the whole stack was thrown away when any error occurred before
    the maximum stack depth was unwound. This is actually a very common
    scenario though. The stacks that got unwound so far are still
    interesting. This removes a large chunk of differences when comparing
    perf script output for libunwind and libdw perf unwinding.

If not, then this could explain the issue you are seeing.

Cheers

-- 
Milian Wolff | milian.wolff@kdab.com | Software Engineer
KDAB (Deutschland) GmbH&Co KG, a KDAB Group company
Tel: +49-30-521325470
KDAB - The Qt Experts